In 2021, Musk announced on Twitter that Tesla would no longer accept Bitcoin payments due to environmental issues, causing the currency price to plummet and sparking people's concerns about the environmental impact of the PoW mining model. The Web3 community still remembers this.

In 2025, Musk's xAI is building what may be the world's largest AI supercomputing cluster, governments are scrambling to legislate to promote AI innovation, and almost no one is questioning its energy consumption anymore.

Whether it is the energy consumption of data centers required for AI or the energy consumption required for the operation and maintenance of the Bitcoin network, the two have begun to compete, and the former is showing a trend of surpassing the latter.

“We have little real insight into the actual power consumption of AI”

A recent peer-reviewed paper published in the scientific journal Joule suggests that by the end of 2025, AI’s electricity consumption could account for 49% of global data center electricity usage, exceeding Bitcoin’s well-known “energy appetite.”

Alex de Vries-Gao, a PhD candidate at Vrije Universiteit Amsterdam and a long-time critic of bitcoin’s energy consumption, found that by January 2026, AI’s power needs could reach 23 gigawatts, or about 201 terawatt hours (TWh) per year. Bitcoin currently consumes about 176 TWh per year.

"Tech giants are well aware of this trend, with companies like Google even referring to a 'power capacity crisis' as they expand data center capacity," he wrote on LinkedIn. "But at the same time, these companies are reluctant to disclose the numbers."

“We have never seen similar growth in energy consumption since ChatGPT ignited the AI craze. This makes it almost impossible to truly grasp the actual power consumption of AI.”

Unlike Bitcoin, which is transparent and can be directly estimated by computing power, AI’s energy consumption is deliberately opaque. Microsoft and Google acknowledged in their 2024 environmental reports that electricity use and carbon emissions have increased significantly, mainly due to AI, but they only disclosed the total consumption of the entire data center without breaking down the specific share of AI.

Since tech companies refuse to disclose AI energy consumption data, de Vries-Gao turned to tracking the chip supply chain. He chose to analyze TSMC's packaging capacity because almost all advanced AI chips require TSMC's technology.

The study found that Nvidia alone will use about 44% and 48% of TSMC's CoWoS (advanced packaging) capacity in 2023 and 2024, respectively. Adding AMD's share, these two companies alone are enough to manufacture AI chips that consume 3.8 gigawatts of electricity, not counting other manufacturers.

de Vries-Gao's model predicts that without adding more production capacity, the total power consumption of AI can reach 23 GW by the end of 2025. TSMC has confirmed that it will double its CoWoS production capacity again in 2025.

Tech giants’ actions

The demand for electricity is unlikely to slow. Nvidia and AMD announced record revenues, and OpenAI announced a $500 billion super data center project called Stargate. In fact, AI has become the most profitable business in the tech industry—the market value of any one tech giant is larger than the entire $3.4 trillion cryptocurrency market. As a result, environmental concerns have been put on hold.

On the one hand, the strong demand for electric energy has forced technology giants to take action to update the way they obtain energy. On the other hand, we have also seen that DePIN is playing an increasingly important role.

On June 3, Meta signed a 20-year agreement with Constellation Energy to purchase nuclear power energy, marking Meta's first official entry into the nuclear energy field to meet the growing power needs of data centers.

Meta will buy about 1.1 GW of energy from Constellation’s Clinton Clean Energy Center in Illinois — the entire output of the plant’s lone reactor — starting in June 2027. The parties said the long-term agreement will support the plant’s continued operations and its license renewal. Without Meta’s commitment, the plant, which has been operating since 2017 on subsidies for zero-emission electricity, would have been at risk of closure.

The deal will also increase the Clinton Nuclear Power Plant’s generating capacity by 30 megawatts, although specific terms have not been disclosed. It’s worth noting that the plant will not directly power Meta’s data centers, but will continue to feed power into the regional grid, while also helping Meta achieve its goal of 100% clean energy.

“We are proud to work with Meta…They recognize that supporting the renewal and expansion of existing nuclear power plants is just as important as developing new energy sources,” said Joe Dominguez, president and CEO of Constellation. “Many times, the first critical step to moving forward is not to look back.”

Google recently pledged to fund the development of three new nuclear power projects, and previously partnered with Kairos Power, a developer of small modular reactors (SMRs). Amazon also invested more than $500 million in the development of SMRs last October, and acquired a data center park powered by the Susquehanna Nuclear Power Plant in March 2024. In March of this year, Amazon, Google, and Meta also jointly signed an initiative initiated by the World Nuclear Association, calling for a tripling of global nuclear power capacity by 2050.

DePIN’s help

DePIN becomes another way to supplement energy consumption by obtaining benefits through blockchain, distributed nodes, and resource sharing.

Taking AI large language models, which require huge computing power, as an example, traditional central processing units (CPUs) are inefficient when processing large-scale parallel computing, so GPUs are usually used to accelerate the training and reasoning process. However, even with GPUs, it is still difficult to meet the computing resource requirements for large language model training and reasoning. AI chips (such as Google's TPU) and quantum computing are both seen as the computing methods required in the AI era.

Training a large model can take days or even weeks and consumes a lot of electricity. It is estimated that the energy consumed by the GPT-3 training process is equivalent to the electricity used by a household for several years. This high energy consumption not only has a negative impact on the environment, but also increases operating costs and raises sustainability issues.

Taking decentralized cloud computing in the DePIN field as an example, by connecting idle computing resources to the cloud computing blockchain network, high-performance, low-latency, and low-cost computing resources are provided to computing power demanders, and profits are obtained by "on-chaining idle computing resources". While fully utilizing computing power resources, computing power resources are fully circulated to participate in the market mechanism, which not only avoids waste but also solves the computing power bottleneck problem.

Conclusion

By the end of 2025, AI may consume nearly half of the electricity in the world's data centers, and this does not include the energy consumption of Bitcoin. It is expected that as AI technology becomes more popular, clean energy and renewable energy will receive more attention and investment. Converting the energy source to nuclear power is one way, and DePIN, which changes the way energy is distributed, is another way. Stay tuned with PowerBeats.